Anda di halaman 1dari 4

Comparison of RBF and MLP Neural Networks

in Short-Term Traffic Flow Forecasting


J. Abdi, Student Member IEEE, B. Moshiri, Senior Member IEEE, and A. Khaki Sedigh

Abstract--Expanding mathematical models and forecasting the


traffic flow is a crucial case in studying the dynamic behaviors of
the traffic systems these days. Artificial Neural Networks (ANNs)
are of the technologies presented recently that can be used in the
intelligent transportation system field. In this paper, two
different algorithms, the Multi-Layer Perceptron (MLP) and the
Radial Basis Function (RBF) have been discussed. In the training
of the ANNs, we use historic data. Then we use ANNs for
forecasting a daily real time short-term traffic flow. The ANNs
are trained by the Back-Propagation (BP) algorithm. The
variable coefficients produced by temporal signals improve the
performance of the BP algorithm. The temporal signals provide a
new method of learning called Temporal Difference BackPropagation (TDBP) learning. We demonstrate the capability
and the performance of the TDBP learning method with the
simulation results. The data of the two lane street I-494 in
Minnesota city are used for this analysis.
Index Terms-- Artificial Neural
Transportation System, Forecasting.

Networks,

Intelligent

I. INTRODUCTION

RAFFIC flow forecasting is an important factor in traffic


control, and route guidance system. Short-term traffic
flow forecasting issue has been presented for at least two
decades and a variety of methods have been developed. In
general, intelligent computational methods that Neural
Networks (NNs) are at the head of them are very common. In
general, these NNs can be categorized into ANN [1], [2], [3],
ANN with parallel BP learning [3], parallel self-scaling quasiNewton neural network [4] and RBF neural networks which
has been introduced in [5] for the first time.
In this paper, the MLP and RBF architectures are provided
in section two and three, respectively. In section four, the
extended learning formula and how they are used for training
neural networks are presented. The criteria formulae for
forecasting methods and the simulation results are provided in
section five and six, respectively. At the end, the conclusion is
summarized in section eight.

J. Abdi is with the Department of Electrical Engineering, Islamic Azad


University - NazarAbad Branch, Tehran, Iran (email: j.abdi@ieee.org).
B. Moshiri is with the School of Electrical and Computer Engineering,
University of Tehran, Tehran, Iran (email: moshiri@ut.ac.ir).
A. Khaki Sedigh is with the Department of Electrical Engineering, K. N.
Toosi University of Technology Tehran, Iran (email: sedigh@kntu.ac.ir).

II. APPLICATION OF MLP NEURAL NETWORKS IN


FORECASTING
In order to approximate any degree of complexity and
without prior knowledge about solving the problem, the ANN
models have received more attentions and are considered as
alternatives for other traffic flow forecasting models [1] [6],
[7], [8]. A MLP with fuzzy logic ramp metering algorithm
forecasts the congested freeway data in [9]. The basis of the
MLP neural network is shown in Fig 1. Mathematically, a
neuron k can be described as follows:
p

y k = (uk k ) where uk = wkj x j

(1)

j =1

and where x j is the value of the jth input neuron, wkj is the
weight between the kth input neuron and the jth hidden layer
neuron, uk is the output of linear combiner, k is the
threshold, (.) is the activation function, and y k is the output
signal of the neuron. In this model, sigmoid function is chosen
to be the activation function for each neuron. The activation
function is presented as follows:
d
1
(2)
( j) =
;
a > 0 and a = w x + b
1 + exp( a )

where

hi i

i =1

whi is updated during training and bn is a bias unit of

the network. In order to test the learning accuracy of the


ANNs, we define an error term as the following:
1 N
2
(3)
E = ( yi y i )
2 i =1
Where y i is the ith real data, y i is the ith forecasted data and

N is the number of samples. The algorithms responsible for


this method is called the learning rule and the most commonly
used one is the BP algorithm rule. The BP algorithm uses the
steepest descent rule for the minimization of the Mean Square
Error (MSE). The BP learning rule is defined by:
E
(4)
+ wkj
wkj (n + 1) = wkj (n ) +
w
kj

where is the learning rate, wkj is the change of weight


during the last iteration, and is the momentum rate. Here,
the recommended values for the simulation are consisting of
momentum ratio is 0.8, learning ratio is 0.7, the number of
epoch in each training is 1000. The data are scaled to vary
between [-1, 1], and the network has two hidden layers.

978-1-4244-8542-0/10/$26.00 2010 IEEE

xi (t n)

xi (t n + 1)

y (t )

y (t )

where each y t is an observation vector available at time t,


e(t )

xi (t )

TD learning which is characterized by its fast computation and


in some cases lower error in comparison with the classical
learning methods [11]. The advantage of TD learning is that it
can update weights incrementally and can converge to a
solution faster than other algorithms [12]. The observation and
the outcome sequence are in the form of y1 , y 2 , y 3 ,..., y m , y ,

1 t m , and y is the outcome of the sequence. For each


observation, the learning agent makes a forecasting from

y,

and then forms the sequence y 1 , y 2 ,..., y m . By assuming the


Fig. 1. The basic structure of MLP NN diagram

III. RBF NEURAL NETWORKS IN FORECASTING


The RBF neural network is in fact a kind of feed-forward
neural network with a single hidden layer. Fig. 2 shows the
basic structure of a RBF networks with n inputs and an output.
RBF neural networks are the supervised neural networks that
seem similar to MLPs [10]. The input and the output of a RBF
is formulized by:

ui = g ( x ci i )

i = 1,..., n

learning agent is an ANN and updating for a weight w of the


network by using the classical gradient descent method the
updating rule is:
m

w = w E = ( y t y ) w y t

(9)

t =1

where w E is the gradient vector. In [11], the incremental


updating rule for (9) is presented as:
t

wt = ( y t y t 1 ) w y k , t = 1,2,..., m , y m +1 = y

(10)

k =1

and
y = Cu .

To highlight recent predictions, an exponential factor is


multiplied to the gradient term:
t

w0

x i (t n )

y (t )

w1

xi (t n + 1)

w2

y (t )

xi (t )

wn

Fig. 2: The basis structure of RBF NN diagram

where x , u , y , C , ci and i > 0 represent the input, the


output of the hidden layer, the network output, weight matrix,
centre and variance of the i th neuron, respectively. The
activation function used here for the h th hidden unit is chosen
Gaussian and the distance is Euclidean that is:
(h ) = exp r 2 h2 . A RBF depends only on the distance

between the input and the centre of the RBF that it is


represented by r. The parameter h indicates the width of
Gaussian function and is updated during training. So the
algorithm is continued until the least mean square (LMS)
condition is reached. The weight can be updated by the
discussed BP algorithm.

wt = ( y t +1 y t ) t k w y k ,

(11)

The (11) is a set of learning rules which is called TD( ) ,


where is a constant value. The has crucial role in
learning as it discussed in [12]. Based on the upper and the
lower band of , when = 0 , (11) will be as similar as (9):

wt = ( y t +1 y t ) w y k

(12)

Based on the new delta rule which is presented in (12), the


BP algorithm can be modified by replacing in (4) and we call
it TDBP learning algorithms.
V. CRITERIA FORMULA FOR COMPARISON METHODS
There are a lot of error indices to display the system states
accuracy's forecasting which are used for evaluating the
algorithm. Based on the illustrated key features of these
criteria and the similarity of them, we found that Mean
Absolute Percentage Error (MAPE), Root MSE (RMSE) and
Variance Absolute Percentage Error (VAPE) were used to
compare in forecasting the traffic flow performance of the
three models as formulated as the following:

MAPE =

IV. NEW LEARNING ALGORITHM: TDL


TD learning is a forecasting method. Based on the
mentioned issue, the BP algorithm is the most common
learning algorithm for ANNs. But the most important and new
learning algorithm based on the Reinforcement Learning is

0 1

k =1

RMSE =

1
N

i =1

y i y i
100%
yi
N

(1 N ) ( y i yi )2

(13)

(14)

i =1

VAPE = Var ( y i yi yi ) 100%

(15)

VI. SIMULATION RESULTS


The proposed algorithm is tested on a real case study to
display its capability to forecast the traffic states. The test of
the proposed traffic state predictor is performed with the real
traffic measurement data obtained from a 2-lane eastbound
stretch of the I-494 freeway in Minnesota, USA. For training
the model, the momentum learning rule with a learning rate of
0.1 is selected for ANNs. The stop criterion for the MLP and
RBF neural networks learning model is that the MSE for the
cross validation data reach a threshold of 0.01, while the
maximum epoch for the supervised learning is limited to 1000.
The data obtained from 1st of January 2009 is used for
training the model. After the ANNs models are trained, they
are applied to the test data set for prediction. The forecasting
outputs are compared with the observed traffic flow data to
test the performance of the two neural network models.
Simulation results of mentioned algorithms are shown in Fig.
3 and Fig. 4.

Fig. 3: A part extracted from the traffic flow forecasting of station 708 with
BPL algorithm

TABLE II
LEARNING TIME COMPARISON OF THE NETWORKS FOR PROPOSED METHODS
WITH BPL AND TDBPL ALGORITHMS

MLP
RBF

BPL (sec.)

TDBPL (sec.)

312.7
261.3

165.1
145.3

VII. CONCLUSIONS
In this paper we presented a new algorithm called TDBP
learning, and ANNs with the learning coefficients for
forecasting the real case study traffic flow. The results show
that in the same situation TDBP learning algorithm is
approximately 2.3 times faster than the BPL algorithm. The
performance criteria algorithm's of this are also approximately
1.6 times better than of the BP learning algorithm. In the BP
learning algorithm, the addition of any new pattern will
influence on the weights of all links. The same weights may
be dragged by different learning patterns in different
directions. As indicated in the table 2, training of the network
will require extreme computational time. In the TDBP
learning algorithm, the new pattern affects on the weights of
all links, too. As illustrated in our algorithm, each step of
forecasting has its past content. Therefore, the weights are not
so far from its actual value when they change. Consequently,
the computational time for training the network will decrease.
By comparing the results shown in Fig. 3 and Fig. 4, TDBP
learning algorithm has more efficiency rather than BP learning
algorithm. Using past information differentials and the time
importance of these information for learning provide an added
virtual memory to system which raise our networks
performance.
VIII. REFERENCES
[1]

[2]
Fig. 4: A part extracted from the traffic flow forecasting of station 708 with
TDBPL algorithm

The comparative results and the learning time for the


illustrated networks are specified in tables 1 and 2. Although
the improvement obtained from the new learning algorithm
appears to be considerable, it should be noted that RBF neural
networks has better results than MLP networks.
TABLE I
THE PERFORMANCE COMPARISON OF MAPE, RSME AND VAPE ERROR
INDICES OF PROPOSED METHODS WITH OF BPL AND TDBPL ALGORITHMS

MLP
RBF

MAPE (BPL)

RSME (BPL)

VAPE (BPL)

6.73
5.34

6.31
5.28

6.59
4.60

[3]
[4]

[5]
[6]
[7]
[8]

MLP
RBF

MAPE
(TDBPL)

RSME
(TDBPL)

VAPE
(TDBPL)

4.23
3.13

5.24
4.19

3.71
2.60

[9]

H. R. Kirby, S. M. Watson and M. S. Dougherty, Should we use neural


network or statistical models for short-term motorway traffic
predicting, International Journal of Predicting, vol. 13, no. 1, pp. 43
50, 1997.
W. Zheng, D. Lee, and Q. Shi, Short-term traffic flow prediction:
Bayesian combined neural network approach, Journal of Trans. Eng.,
vol. 132, no. 2, pp. 114121, 2006.
G. Tan, Q. Deng, Z. Tian, and J. Yang, A Dish Parallel BP for Traffic
Flow Forecasting, Inter. Conf. on Computational Intelligence and
Security (CIS 2007), pp.546-549, 2007.
G. Tan, H. Shi, F. Wang and C. Deng, Short-Term Traffic Flow
Forecasting Based on Parallel Quasi-Newton Neural Network,
International Conference on Measuring Technology and Mechatronics
Automation, ICMTMA '09, vol. 3, pp. 305-308, 2009.
M. S. Ahmed and A. R. Cook, Analysis of freeway traffic time-series
data by using Box-Jenkins techniques, Transportation Research
Record, no. 722, pp. 1-9, 1979.
M. S. Dougherty and M. R. Cobbett, Short-term inter-urban traffic
forecasts using neural networks, International Journal of Forecasting,
vol. 13, pp. 21-31, 1997.
L. Florio and L. Mussone,. Neural network models for classification
and forecasting of freeway traffic flow stability, Control Engineering
Practice, vol. 4, pp. 153-164, 1996.
E. I. Vlahogianni, M. G. Karlaftis and J. C. Golias, Optimized and
meta-optimized neural networks for short-term traffic flow prediction: A
genetic approach, Transportation Research Part C, vol. 13, pp. 211234, 2005.
C. Taylor and D. Meldrum, Freeway traffic data forecasting via
artificial neural networks for use in a fuzzy logic ramp metering
algorithm, Intell. Vehicles Symp., pp. 308-313, 1994.

4
[10] H. B. Celikoglu and H. K. Cigizoglu, Modelling public transport trips
by radial basis function neural networks", Mathematical and Computer
Modelling, vol. 45, no. 3-4, pp. 480-489.
[11] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction,
MIT Press, Cambridge, Mass, 1998.

[12] J. Abdi, M. Fatourechi, C. Lucas, A. Khaki Sedigh and G. H.


FamilKhalili, "Control of Multivariable Systems based on Emotional
Temporal Difference Learning Controller", International Journal of
Engineering, National Centre for Science Research, vol. 17, no. 4, pp.
363-376, 2004.

Anda mungkin juga menyukai